All models are wrong but all can be useful: Robust policy design using prediction pools

We study the design of monetary policy rules robust to model uncertainty using a novel methodology. In our application, policymakers choose the optimal rule by attaching weights to a set of well-established DSGE models with varied financial frictions. The novelty of our methodology is to compute each model's weight based on their relative forecasting performance. Our results highlight the superiority of predictive pools over Bayesian model averaging and the need to combine models when none can be deemed as the true data generating process. In addition, we find that the optimal across-model robust policy rule exhibits attenuation, and nests a price level rule which has good robustness properties. Therefore, the application of our methodology offers a new rationale for price-level rules, namely the presence of uncertainty over the nature of financial frictions. • We study the design of simple monetary policy rules robust to model uncertainty. • Predictive pooling weights models based on their relative forecasting performance. • The models used in our application differ over the nature of financial frictions. • Results highlight the superiority of predictive pools over Bayesian model averaging. • The optimal robust rule exhibits attenuation and is close to price-level targeting.